I was thinking earlier about “voxel 3-D printing” with spot-welded bearing balls, where each new ball added to the workpiece is located by contacts with three existing balls, then spot-welded to each of them in turn. That way, the end effector doesn’t need to have extremely precise location abilities, because the location is in the precision of the bearing balls, which can easily and very inexpensively be submicron.
However, spot-welding them will both induce stresses on the structure as the nugget expands and then contracts, and also reduces the precision of the distance between the centers of the balls; the bigger the spot welds, the bigger this effect. And with any practical size of spot weld, the resulting structure will be much weaker than . Also, it requires high power input from the end effector, even though it’s electrical power.
The standard mechanical-engineering solution to locating and fastening parts accurately is to use locator pins (or other features) separate from the fasteners, such as screws. By chamfering or tapering locator features, it becomes possible to assemble parts with greater precision than the precision of the manipulators, in the same way that the balls in my first paragraph would provide high precision. Somewhat analogously, the position of a lathe saddle can be indicated much more precisely by a dial indicator (or, nowadays, a digital readout) than by the graduations on the handwheel, because (aside from backlash) the kinematic chain of the dial indicator does not have to bear the load of feeding the tool into the work.
The wedging action of a tusk-tenon joint makes such a permanent joint for a somewhat related reason: the load on the joint is orthogonal to the direction of movement of the wedge, so it does not tend to dislodge it.
So I think that perhaps the best way to assemble things for a permanent, rigid mechanical connection is:
Position them in a precise place using positioning features such as a Maxwell kinematic coupling.
Hold them in that place using a fastening system that can handle all the variations in position that the positioning system can produce, without producing large enough loads during the holding operation to create positioning errors. For example, two parallel plates sliding against one another are a planar joint, with three degrees of freedom, until one or more screws through oversize holes in one into tapped holes in the other add enough friction to prevent movement. A spherical ball-and-socket joint also has three degrees of freedom until enough friction is similarly added. With a serial kinematic chain of three joints (of two or three degrees of freedom), you can provide all six degrees of freedom; putting them close together and putting more than one such chain in parallel can provide greater rigidity. (There might be a way to do it with just two joints, but I can’t see it.)
Lock the holding/fastening mechanism with something adequately permanent, like self-propagating high-temperature synthesis to fuse parts together, some safety lockwire, a jam nut, or just a circlip or similar spring. The loads will be borne by the holding/fastening mechanism, not by the positioning mechanism or by the locking mechanism, because the locking mechanism only serves to prevent the fastening mechanism from coming unfastened.
These three functions are not always so independent; in a four-jaw lathe chuck, for example, each jaw fulfills both the positioning function (when the other jaw is far away) and the holding function (when it’s adding pressure to the part and thus friction to both the part and the other jaw). But I think separating them will generally improve precision. At times, in a lathe chuck, the moving function in two rotational degrees of freedom is provided by tapping the workpiece up against the flat face of the chuck, before holding the part in place by tightening the jaws.
You could perhaps drench the balls in a viscous liquid that later forms a glass, which can perhaps later be annealed into a glass-ceramic, so that they are positioned by the precise Hertzian contact between the balls, but then held in place by the glass or glass-ceramic matrix. This will work best if the matrix is nearly as hard as the balls (in the sense of Young’s modulus) or even harder. In effect, the matrix foam is the real object; the balls are just there to provide it with precise dimensions, and they could be hollow bubbles or even removed entirely after the matrix hardens. Hollow fused-quartz bubbles would probably be especially useful for this purpose.